Probably, but not one you own. AI models do thousands of small calculations at once, which is exactly what a graphics card is built for and exactly what a normal processor is bad at. You can borrow one by the hour for the price of a coffee instead of buying a $1,600 card.
If you have arrived here from a search like "do I need a GPU for AI" then you have probably already had the experience that prompts the question. Something you wanted to try said it needed a graphics card. Or it ran, but so slowly that you assumed it had crashed. Or a forum told you to buy a $1,600 card and you closed the tab.
The honest answer is yes, you need one, and no, you almost certainly should not buy it. Here is what a GPU actually does, why this particular kind of work needs it, and the three ways to get access to one without owning anything.
What a GPU actually does that a processor cannot
Your computer already has a processor — the CPU — and it is very good. It just happens to be good at the wrong thing.
A CPU has a small number of extremely fast cores, typically eight or sixteen. Each one works through instructions in order, very quickly. It is a brilliant generalist, which is why it runs your operating system, your browser and your spreadsheet.
A GPU — a graphics processing unit — has thousands of cores that are individually slower and all work at the same time. It was built to colour in millions of pixels simultaneously for video games, because every pixel can be worked out independently of every other pixel.
Physically, a GPU is a card: a board about the size of a paperback that slots into a computer, which is why people who work with them just say the card. The machine it slots into, when it lives in a data center instead of under a desk, is called a GPU server. Those three words — GPU, graphics card, the card — all mean the same object, and this guide uses them interchangeably because the rest of the internet does.
Here is the useful analogy. A CPU is one very fast mathematician. A GPU is a stadium full of school children who can each do simple sums. Ask for a complicated proof and the mathematician wins easily. Ask for two million simple additions and the stadium finishes before the mathematician has warmed up.
Why AI work specifically needs one
Running an AI model is, underneath everything, multiplying enormous grids of numbers together. Not clever mathematics — repetitive mathematics, in vast quantity, where every individual sum is independent of the others.
That is the stadium's problem, not the mathematician's. It is the same shape of work as colouring in pixels, which is a genuine historical accident: the hardware built for video games turned out to be the hardware AI needed, and the whole field took off partly because the equipment was already sitting in people's computers.
The difference is not a small percentage. Generating one image takes about eight seconds on a rented RTX 4090. On a modern laptop processor the same image takes six to eight minutes. A batch of 500 images is a lunch break on a graphics card and roughly three days on a processor.
See the numbers
| Model | Card | Time | Cost |
|---|---|---|---|
| SDXL Turbo | RTX 3090 | 10 min | $0.04 |
| Stable Diffusion XL | RTX 4090 | 67 min | $0.50 |
| FLUX.1 dev | L40S | 150 min | $2.23 |
What happens if you try without one
It usually runs. That is the confusing part. There is rarely an error message that says "you need a GPU" — instead the thing starts, appears to work, and then takes so long that you assume it has hung.
Three symptoms tell you this is what is happening:
- A progress bar that moves, but would take hours to finish
- Your computer's fans running flat out while everything else becomes sluggish
- A chat model producing about one word a second, like a bad phone connection
None of that is broken. It is a processor doing a graphics card's job, honestly and slowly.
What is VRAM, and how much do you need
VRAM is the memory on the graphics card itself — think of it as the desk the model has to spread its work out on. Too small a desk, and the job just will not start.
This catches everyone out once, because it does not behave like speed. If a card is slow, your job takes longer. If a card is short on memory, your job does not happen at all. There is no partial credit.
The rough guide for chat models is one gigabyte per billion parameters, once the model has been compressed for everyday use. An 8-billion-parameter model needs about 8GB. A 70-billion one needs about 48GB, which is more memory than any consumer graphics card has ever shipped with.
For image work, 12GB is the floor and 24GB is where it stops being annoying. For video, or for training a model on your own material, start at 48GB.
Can you run AI on a laptop?
Some of it, genuinely well.
A recent MacBook is better at this than people expect, because Apple's chips share one pool of memory between the processor and the graphics part. A MacBook with 32GB of memory can run chat models that would need a $2,000 desktop card to run on Windows. It will not be fast, but it works, and it is free.
A Windows laptop with a dedicated graphics card of 8GB or more will run small chat models and will run image generation, slowly. A laptop without a dedicated card will technically run small models and will not enjoy it.
What no laptop does is the large end: the 70-billion-parameter models, video generation, or fine-tuning anything. Those need memory that does not exist in portable machines, and no amount of patience substitutes for it.
If a laptop covers what you need, use the laptop. We would genuinely rather tell you that than sell you an hour of something you already own.
What can you actually run yourself?
More than most people assume, and the list gets longer every few months. Almost everything worth using is open, free to download, and runs on hardware you can rent for the price of a sandwich.
Chat and writing. Llama, Mistral, Qwen, Gemma and DeepSeek all publish open models. The small ones — around 8 billion parameters — are genuinely useful for summarising, drafting and answering questions about your own documents, and they run in about 8GB of memory. The large ones, around 70 billion, are close enough to the commercial services that most people cannot tell in normal use, and they need about 48GB.
Images. Stable Diffusion XL and FLUX.1 cover nearly everything people want: product shots, illustration, concept art, editing existing photographs. This is the most mature corner of the field and the easiest place to start.
Speech. Whisper transcribes audio better than most paid services and costs a few cents an hour of audio. If you have a backlog of interviews, recordings or podcast episodes, this is usually the first thing people run and the thing that converts them.
Video. Upscaling, frame interpolation, background removal and short generated clips all work now. Video is the hungriest category by a distance, so this is where rented hardware stops being a convenience and starts being the only practical option.
3D and rendering. Blender renders, photogrammetry, and models that turn a photograph into a 3D mesh.
What you cannot sensibly run yourself is the very frontier: the largest commercial models are not published, and no amount of hardware will change that. For most work that gap matters far less than the marketing implies.
If I already use ChatGPT, why would I need any of this?
A fair question, and for a lot of people the honest answer is that you would not.
What you are paying for with a commercial chat service is someone else's hardware, someone else's model, and a monthly fee that does not move whether you use it once or a thousand times. That is a good deal for casual use and it needs no explanation from us.
Three things push people over to running models themselves.
Your data stays yours. If you are working with client material, patient records, unreleased products or anything covered by a contract, sending it to a third-party service is often simply not allowed. A model running on a machine you rented, that is destroyed afterwards, is a different legal proposition.
Volume changes the maths. A subscription is excellent value for a few hundred questions a month and terrible value for processing forty thousand support tickets. At volume, per-hour compute is dramatically cheaper than per-seat or per-token pricing.
You want a model that does one thing your way. Training a model on your own photographs, your own writing style or your own product catalogue is not something a general service will do for you. That is the whole point of fine-tuning, and it needs hardware.
If none of those apply to you, keep the subscription. We mean that.
What is a GPU cluster, and do you need one?
You will see the phrase and it sounds like something you are missing out on. You are not.
A GPU cluster is many machines with many graphics cards, wired together with very fast networking so they can work on one enormous problem at the same time. Clusters exist to train models from scratch — the kind of job that runs for weeks across hundreds of cards and costs millions.
Running a model that somebody else has already trained is a completely different activity. It happens on one card, usually in seconds. Fine-tuning an existing model on your own material sits in the middle and still usually fits on one card, or a machine with a few of them.
So unless you are training a foundation model, which you would know about, one card is the answer. Anyone selling you a cluster for inference work is selling you a cluster.
Does generating AI images damage your GPU?
No, and this question comes up often enough to deserve a straight answer.
Generating images runs the card at full load and high temperature for the length of the job. So does any demanding game, for hours at a time, which is precisely what these cards were designed and warrantied for. The card will slow itself down before it reaches a temperature that would hurt it.
What actually shortens a graphics card's life is heat it cannot escape: a dusty machine, a case with no airflow, a card wedged against a wall. Fix the cooling and the workload is a non-issue. Running a card hard is not the same as running it badly.
The three ways to get a GPU without buying one
Use what you already have. If your machine can do the job, do the job. Free is a very good price.
Use a free cloud notebook. Google Colab gives you a GPU for nothing, with two catches: it disconnects when it feels like it, often in the middle of a long job, and the free tier gets whatever hardware is spare. Fine for learning. Genuinely frustrating for anything you need to finish.
Rent one by the hour. This is what we do. You pick the job, we start a machine with the model already loaded, and you get a link in your browser about 90 seconds later. An RTX 4090 is $0.45 an hour, billed by the minute, and stops billing when you stop working.
The reason renting suits most people is not the hourly rate, it is the shape of the demand. Almost nobody needs a GPU steadily. They need one badly for nine days and then not at all for two months. Owning hardware for that pattern means paying for two months of a card sitting in the dark.
The full arithmetic — including the point where it flips and you genuinely should buy — is in our guide to how GPU rental works and the 20-hour rule. The short version: buying wins after about 4,183 hours of real use, which is 20 hours a week for four years.
Does running AI models cost money?
The models are mostly free. Stable Diffusion, Llama, Whisper and most of what people actually use are open and cost nothing to download.
What costs money is the machine underneath. That is true whether you buy the machine, rent it, or use somebody's API — an API is just renting compute with the hardware hidden and a markup applied.
To put a real number on it: transcribing a one-hour podcast with Whisper costs a few cents. Generating 500 product photos costs about fifty cents. Getting a 70-billion-parameter model to read a long document and summarise it costs a few cents. The costs that surprise people are almost never the jobs, they are the machine left switched on afterwards.
Where to start
Do not start by choosing hardware. Start with the thing you want to make, and let the requirement fall out of that.
Or look through the kinds of work people do here and find the one that resembles yours.